Message ID | 20160823184724.GA11707@p183.telecom.by (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Dne 23.8.2016 v 20:47 Alexey Dobriyan napsal(a): > Do you think kernel build is 100% dominated by gcc? You are wrong! > One small utility called "fixdep" consistently manages to sneak into > profile's first page (unless you have small monitor of course). > > The choke point is this clever code: > > for (; m < end; m++) { > if (*m == INT_CONF) { p = (char *) m ; goto conf; } > if (*m == INT_ONFI) { p = (char *) m-1; goto conf; } > if (*m == INT_NFIG) { p = (char *) m-2; goto conf; } > if (*m == INT_FIG_) { p = (char *) m-3; goto conf; } > > 4 branches per 4 characters is not fast. > > Use strstr(3), so that SSE2 etc can be used. It should be noted that the previous code also matched nested CONFIG_* strings, like CONFIG_PROC inside CONFIG_IKCONFIG_PROC. But these are pointless, so your patch is removing a few false matches. > With this patch, fixdep is so deep at the bottom, it is hard to find it. Nice. > - if (*m == INT_CONF) { p = (char *) m ; goto conf; } > - if (*m == INT_ONFI) { p = (char *) m-1; goto conf; } > - if (*m == INT_NFIG) { p = (char *) m-2; goto conf; } > - if (*m == INT_FIG_) { p = (char *) m-3; goto conf; } Please also remove the INT_* definitions and the test at startup. > + map = malloc(st.st_size + 1); > + if (!map) { > + perror("fixdep: malloc"); > close(fd); > return; > } > + if (read(fd, map, st.st_size) != st.st_size) { > + perror("fixdep: read"); > + close(fd); > + return; > + } I wanted to suggest memmem() instead of strstr() to avoid the extra copy, but that would make use depend on glibc (this _might_ have been the reason to use the manual string search, in fact). So let's leave this as. Michal -- To unsubscribe from this list: send the line "unsubscribe linux-kbuild" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Tue, Aug 23, 2016 at 09:47:24PM +0300, Alexey Dobriyan wrote: > Do you think kernel build is 100% dominated by gcc? You are wrong! > One small utility called "fixdep" consistently manages to sneak into > profile's first page (unless you have small monitor of course). > > The choke point is this clever code: > > for (; m < end; m++) { > if (*m == INT_CONF) { p = (char *) m ; goto conf; } > if (*m == INT_ONFI) { p = (char *) m-1; goto conf; } > if (*m == INT_NFIG) { p = (char *) m-2; goto conf; } > if (*m == INT_FIG_) { p = (char *) m-3; goto conf; } > > 4 branches per 4 characters is not fast. > > Use strstr(3), so that SSE2 etc can be used. > > With this patch, fixdep is so deep at the bottom, it is hard to find it. > > Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> > --- > > scripts/basic/fixdep.c | 67 ++++++++++++++++++++----------------------------- > 1 file changed, 28 insertions(+), 39 deletions(-) > > --- a/scripts/basic/fixdep.c > +++ b/scripts/basic/fixdep.c > @@ -82,8 +82,7 @@ > * to date before even starting the recursive build, so it's too late > * at this point anyway. > * > - * The algorithm to grep for "CONFIG_..." is bit unusual, but should > - * be fast ;-) We don't even try to really parse the header files, but > + * We don't even try to really parse the header files, but > * merely grep, i.e. if CONFIG_FOO is mentioned in a comment, it will > * be picked up as well. It's not a problem with respect to > * correctness, since that can only give too many dependencies, thus > @@ -241,37 +240,22 @@ static void use_config(const char *m, int slen) > print_config(m, slen); > } > > -static void parse_config_file(const char *map, size_t len) > +static void parse_config_file(const char *p) > { > - const int *end = (const int *) (map + len); > - /* start at +1, so that p can never be < map */ > - const int *m = (const int *) map + 1; > - const char *p, *q; > - > - for (; m < end; m++) { > - if (*m == INT_CONF) { p = (char *) m ; goto conf; } > - if (*m == INT_ONFI) { p = (char *) m-1; goto conf; } > - if (*m == INT_NFIG) { p = (char *) m-2; goto conf; } > - if (*m == INT_FIG_) { p = (char *) m-3; goto conf; } > - continue; > - conf: > - if (p > map + len - 7) > - continue; > - if (memcmp(p, "CONFIG_", 7)) > - continue; > + const char *q, *r; > + > + while ((p = strstr(p, "CONFIG_"))) { > p += 7; > - for (q = p; q < map + len; q++) { > - if (!(isalnum(*q) || *q == '_')) > - goto found; > - } > - continue; > - > - found: > - if (!memcmp(q - 7, "_MODULE", 7)) > - q -= 7; > - if (q - p < 0) > - continue; > - use_config(p, q - p); > + q = p; > + while (*q && (isalnum(*q) || *q == '_')) > + q++; > + if (memcmp(q - 7, "_MODULE", 7) == 0) > + r = q - 7; > + else > + r = q; > + if (r > p) > + use_config(p, r - p); > + p = q; > } > } Faster and simpler - good. > > @@ -291,7 +275,7 @@ static void do_config_file(const char *filename) > { > struct stat st; > int fd; > - void *map; > + char *map; > > fd = open(filename, O_RDONLY); > if (fd < 0) { > @@ -308,18 +292,23 @@ static void do_config_file(const char *filename) > close(fd); > return; > } > - map = mmap(NULL, st.st_size, PROT_READ, MAP_PRIVATE, fd, 0); > - if ((long) map == -1) { > - perror("fixdep: mmap"); > + map = malloc(st.st_size + 1); > + if (!map) { > + perror("fixdep: malloc"); > close(fd); > return; > } > + if (read(fd, map, st.st_size) != st.st_size) { > + perror("fixdep: read"); > + close(fd); > + return; > + } > + map[st.st_size] = '\0'; > + close(fd); > > - parse_config_file(map, st.st_size); > - > - munmap(map, st.st_size); > + parse_config_file(map); > > - close(fd); > + free(map); > } This change is not described in the changelog.. What is the rationale? Sam -- To unsubscribe from this list: send the line "unsubscribe linux-kbuild" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, Aug 24, 2016 at 07:01:08AM +0200, Sam Ravnborg wrote: > > - map = mmap(NULL, st.st_size, PROT_READ, MAP_PRIVATE, fd, 0); > > - if ((long) map == -1) { > > - perror("fixdep: mmap"); > > + map = malloc(st.st_size + 1); > > + if (!map) { > > + perror("fixdep: malloc"); > > close(fd); > > return; > > } > > + if (read(fd, map, st.st_size) != st.st_size) { > > + perror("fixdep: read"); > > + close(fd); > > + return; > > + } > > + map[st.st_size] = '\0'; > > + close(fd); > > > > - parse_config_file(map, st.st_size); > > - > > - munmap(map, st.st_size); > > + parse_config_file(map); > > > > - close(fd); > > + free(map); > > } > This change is not described in the changelog.. > What is the rationale? There is no raw memory equivalent of strstr(), so a file is made into a string. -- To unsubscribe from this list: send the line "unsubscribe linux-kbuild" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
--- a/scripts/basic/fixdep.c +++ b/scripts/basic/fixdep.c @@ -82,8 +82,7 @@ * to date before even starting the recursive build, so it's too late * at this point anyway. * - * The algorithm to grep for "CONFIG_..." is bit unusual, but should - * be fast ;-) We don't even try to really parse the header files, but + * We don't even try to really parse the header files, but * merely grep, i.e. if CONFIG_FOO is mentioned in a comment, it will * be picked up as well. It's not a problem with respect to * correctness, since that can only give too many dependencies, thus @@ -241,37 +240,22 @@ static void use_config(const char *m, int slen) print_config(m, slen); } -static void parse_config_file(const char *map, size_t len) +static void parse_config_file(const char *p) { - const int *end = (const int *) (map + len); - /* start at +1, so that p can never be < map */ - const int *m = (const int *) map + 1; - const char *p, *q; - - for (; m < end; m++) { - if (*m == INT_CONF) { p = (char *) m ; goto conf; } - if (*m == INT_ONFI) { p = (char *) m-1; goto conf; } - if (*m == INT_NFIG) { p = (char *) m-2; goto conf; } - if (*m == INT_FIG_) { p = (char *) m-3; goto conf; } - continue; - conf: - if (p > map + len - 7) - continue; - if (memcmp(p, "CONFIG_", 7)) - continue; + const char *q, *r; + + while ((p = strstr(p, "CONFIG_"))) { p += 7; - for (q = p; q < map + len; q++) { - if (!(isalnum(*q) || *q == '_')) - goto found; - } - continue; - - found: - if (!memcmp(q - 7, "_MODULE", 7)) - q -= 7; - if (q - p < 0) - continue; - use_config(p, q - p); + q = p; + while (*q && (isalnum(*q) || *q == '_')) + q++; + if (memcmp(q - 7, "_MODULE", 7) == 0) + r = q - 7; + else + r = q; + if (r > p) + use_config(p, r - p); + p = q; } } @@ -291,7 +275,7 @@ static void do_config_file(const char *filename) { struct stat st; int fd; - void *map; + char *map; fd = open(filename, O_RDONLY); if (fd < 0) { @@ -308,18 +292,23 @@ static void do_config_file(const char *filename) close(fd); return; } - map = mmap(NULL, st.st_size, PROT_READ, MAP_PRIVATE, fd, 0); - if ((long) map == -1) { - perror("fixdep: mmap"); + map = malloc(st.st_size + 1); + if (!map) { + perror("fixdep: malloc"); close(fd); return; } + if (read(fd, map, st.st_size) != st.st_size) { + perror("fixdep: read"); + close(fd); + return; + } + map[st.st_size] = '\0'; + close(fd); - parse_config_file(map, st.st_size); - - munmap(map, st.st_size); + parse_config_file(map); - close(fd); + free(map); } /*
Do you think kernel build is 100% dominated by gcc? You are wrong! One small utility called "fixdep" consistently manages to sneak into profile's first page (unless you have small monitor of course). The choke point is this clever code: for (; m < end; m++) { if (*m == INT_CONF) { p = (char *) m ; goto conf; } if (*m == INT_ONFI) { p = (char *) m-1; goto conf; } if (*m == INT_NFIG) { p = (char *) m-2; goto conf; } if (*m == INT_FIG_) { p = (char *) m-3; goto conf; } 4 branches per 4 characters is not fast. Use strstr(3), so that SSE2 etc can be used. With this patch, fixdep is so deep at the bottom, it is hard to find it. Signed-off-by: Alexey Dobriyan <adobriyan@gmail.com> --- scripts/basic/fixdep.c | 67 ++++++++++++++++++++----------------------------- 1 file changed, 28 insertions(+), 39 deletions(-) -- To unsubscribe from this list: send the line "unsubscribe linux-kbuild" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html